How I Turned Claude Cowork Into the Operating System for My Work
Article2,896 words35 likes
Full build of an agentic work system on Claude Cowork: overnight research desk, morning and evening briefs, zero-thought capture and Notion as system of record. Design rules: buy do not build, scripts return data only, Claude proposes and he decides.
Key claims
Eleven scheduled runs across three Cowork projects
Twelve Notion databases created through the connector
Eight small Python scripts for data without connectors
How I Turned Claude Cowork Into the Operating System for My Work
I just 10x'd my productivity with an agentic system built around @claudeai Cowork. Nothing fancy under the hood, just the tools I already use wired together properly.
You can do it too, so here is what I built:
One place where all my inputs land. X, Telegram, three mailboxes, calendar, meeting notes, the links I save, the tokens I watch, the stuff I hold. Claude reads all of it so I don't have to open eight apps to know what's going on.
An overnight research desk. While I sleep it scans the narratives I track on X, checks what actually happened, pulls prices and onchain flows, digests my Telegram and files whatever I dropped during the day. I wake up to a finished intel page.
A morning brief and an evening brief, delivered where I actually read things: Telegram, plus Notion and email for the full version.
A weekly alpha piece on Mondays, a content plan on Fridays, a Sunday review that plans my week and books deep-work blocks into the calendar.
Capture that costs zero thought. vault:, watch:, idea:, todo:, deadline:, pos: in any Claude chat, in my Telegram bot or as an email subject. A bare link is enough to store alpha I come across without it vanishing into the abyss.
On-demand runs from plain chat: "deep dive on X", "sweep perp DEXs", "catch me up on Telegram", "draft a Hyperliquid piece" and much more.
Notion as the single system of record. Twelve databases Claude reads and writes every run, from Radar and Watchlist to Content calendar, Research, Wallets and Positions.
A Gmail hub. All my mailboxes forward into one, everything labelled on arrival, so Claude reads one inbox and knows what's admin, what's finance, what's a newsletter and what needs a reply.
Typefully as the source of truth for X. Claude can see what I'm working on, can draft into it, help me research and more. Meanwhile I use Todoist for tasks, and Fireflies for meetings. All existing tools, I already used but never was holistically able to keep track of. Now everything is in one place.
Eight small Python scripts underneath for the data no connector provides: X via Grok, CoinGecko, DexScreener, Perplexity, Telegram, Typefully, and wallet balances from explorers.
Why?
Working in crypto is messy, and you probably know that too.
Every day I have to manage thousands of things. I write content for my own account, manage my portfolio, trade, do things onchain, try to keep track of all the opportunities and alpha flying around, maintain a network, and talk to people. Then there is the A1 side: the business, conversations with clients, research production, publishing. All of that happens across X, Telegram, email, Notion, Google Docs, a few dashboards and whatever tab I forgot to close. And the market doesn't stop, so none of it has office hours, making things even harder to manage.
So the problem became evident. It was never a lack of information or tooling. It was that everything lived in a different place and nothing talked to each other. Todos in my head, links in five chats, a watchlist in browser tabs, overflowing bookmarks tab on X that I never revisit, positions spread across wallets and chains, a Notion I had stopped opening bc it got too messy, and so on.
To make it concrete: before, the first hour of every day was catching up. Scroll X, scroll Telegram, open three inboxes, check the charts, remember what I promised to whom, and only then start working. Now there is one page waiting at 07:15 that already did all of that, and I start the day with the actual work.
What I needed was a connective layer. Something that sits across all of these tools, pulls the threads together, runs the boring parts on a schedule and leaves me with one page to read in the morning and one place to drop things during the day.
So I experimented with a lot of tooling, models and harnesses. Used built-in AI tools in Google and Notion, played around with everything from OpenClaw to Hermes, used models from Claude or OpenAI to Kimi and GLM, ran local models on my DGX Spark, and so on.
I learned a lot along the way and what the right fit is for you depends on your use case. But for my specific case here, Claude Cowork is just the easiest solution. A lot of stuff (connectors for Gmail, Calendar, Drive, Notion, Todoist, Fireflies) comes out of the box, scheduled tasks run in the cloud so nothing has to stay on, maintenance overhead is low, and there is much less friction than in a system you build yourself from the ground up. I know because I tried.
How It's Wired
The mental model is simple. Claude does the reading and the thinking. Small scripts pull data when needed. Notion holds state, so every run starts from the same place. Output goes to Telegram because that's where I am most reachable anyway.
A few things I decided early and haven't regretted:
Buy, don't build. Already explained above, but ICYMI: Cowork already has connectors for Gmail, Calendar, Drive, Notion, Todoist and Fireflies. So I only had to write code where there is no connector.
Scripts never summarize, they only return data. The moment a script starts interpreting, the model loses the raw material it needs to think, and using Claude as the brain was also a deliberate choice based on my preference on what powers the reasoning.
Nothing gets replaced. Typefully, Todoist, Notion, Fireflies all stay. Claude works with them, not instead of them and I didn't have to build everything from scratch, just connect it.
Claude proposes, I decide. It suggests tokens for the watchlist, deep dives to run, content angles, status changes. It doesn't act on any of that without me (and also does not manage my wallets).
A couple of things can't run in the cloud sandbox, so a small job on my Mac handles those and syncs through Drive. That's the entire infrastructure.
Zero HQ (Notion)
One private page with twelve databases is the backbone of pretty much everything. I gave Claude a build prompt with the schema and it created the whole thing through the Notion connector in one session. I didn't set up a single database by hand, which is the only reason I'm using Notion again.
The two that drive the research side are Radar and Watchlist. Radar is the list of narratives I care about right now, things like perp DEXs, Hyperliquid and its ecosystem, crypto x AI and robotics, privacy tech, prediction markets, regulation, etc. Each row has search queries, a curated X list of accounts I trust on that topic, a cadence and a Heat field. The overnight run reads Radar to know what to scan and which accounts to weigh, and can propose new narratives when something keeps showing up. Watchlist covers assets with a thesis, an invalidation and a status. Claude updates price, sentiment and the nightly signal, but it never touches the thesis. Only I do that, usually straight from a conversation with Claude. If something hits the invalidation text, the alert flips on.
Vault is my reading list and alpha collection. Every link I drop anywhere, to my TG bot, in a Claude convo or in a mail I send myself, ends up there, gets read and summarized overnight, and the best ones show up in the morning brief.
Obviously I can't fade Telegram. TG Chats is one database with one row per chat, labelled and with topic tags. That alone massively helps to stay on top of things, and the tags let me ask stuff like "what did the agent channels say about onchain payments this week" in any Claude convo (or on the bot) without doomscrolling, and have tasks around communication added or resolved automatically.
Wallets and Positions are the portfolio side of things. I don't want to automate the management and Claude can't execute anything. The idea was just to build something that tracks everything, no matter if or where an asset is listed or what chain it's on. So I collected and entered every address and account once, and now the nightly snapshot fills balances and prices across all chains, NFTs included, while a few manual rows cover the things no API knows about (pre-TGE investments, for example). Each position relates to its Watchlist asset, which ties my research and my theses together with my actual holdings in one place.
The rest is plumbing that runs itself. Content calendar mirrors Typefully. Research holds the deep dives. Briefs is where every run writes its output, so I can always go back and see what Claude did and why. Meetings turns Fireflies transcripts into decisions and tasks fully automatically, Projects gets flagged when progress stalls, and Admin & Deadlines catches the admin stuff I tend to forget about.
The Daily Rhythm
At the system's core are eleven scheduled runs across three Cowork projects (Ops, Research, Content). Each project has its own instructions and the scripts it needs. Six run on a schedule, five on demand.
The morning brief at 07:15 is the one that I start my day with. Top 5 for today, calendar with a prep line per meeting, inbox by category, tasks from Todoist, yesterday's meetings listed and recapped, the content queue outlined and pending items flagged, breaking down everything on markets and narratives (helping me to find inspiration for posts or starting points for research), updates me on my portfolio and watchlist, recommends readings, reminds me of Telegram replies I owe, deadlines I'm about to miss, and much more (all in <1000 words). And in case I'm on the road or otherwise afk, I get a shorter version with a link to the full page delivered to my Telegram.
The evening brief at 18:30 closes out the day and recaps meetings filed, captures processed, dates that (e.g. from mail or elsewhere) turned into deadlines, etc., giving me an outlook for what's at hand tomorrow.
Monday 06:30 is the time for the weekly alpha brief. Or more specifically: a narrative map with heat trend and a contrarian angle per vertical, five numbers that tell the story of the week, onchain flows, watchlist review against thesis and invalidation, portfolio review, top Vault (aka reading list) items, three potential research deep dives to run (which Claude will do along a pre-defined workflow, leveraging tools that make outputs much stronger than any generic chat window query), and three potential content angles (aligning with what ppl are actually talking about on CT, what topics have increasing social velocity, etc.).
The five on-demand ones I fire from any chat. "deep dive on X" runs the full research workflow and writes a Research page. "sweep perp DEXs" does a fast pass across one Radar list when something is moving. "vault digest" clears the reading backlog. "draft X" produces a Typefully draft. "catch me up on Telegram" gives me the reply queue and what I missed.
The Overnight Run
Powering most of the insights for my morning briefings is one heavy research run that runs at 05:30, every day. It's the main data feed of the entire system and is being run from the "Research" project in Cowork.
Capture. Reads my bot chat since last time and routes everything with a prefix: vault, watchlist, content idea, todo, deadline, position.
Telegram digest. Deep chats: decisions, open questions to me, action items. Skim chats: links into the Vault. Only what's labelled, and only the digest leaving Telegram.
Radar scan. Per narrative: x_scan.py hot across the curated list for what X is talking about, pplx(.)py answer for what actually happened, a web search as a third source. Then Claude reconciles them. Heat goes up or down.
Watchlist scan. cg(.)py for listed tokens, dex(.)py for onchain-only ones with a liquidity check, x_scan.py asset for sentiment, Perplexity for news. One line per asset, Alert if the 7d move is above 25 percent or something otherwise material happened.
Portfolio snapshot. folio(.)py snapshot walks every wallet (Helius, Etherscan, Hyperliquid, etc.), prices everything, reprices the manual rows if needed, and flags anything that moved more than 10 percent overnight.
Attention radar. DexScreener boosted pairs and new profiles, CoinGecko trending on HyperEVM, Solana, RH Chain and Base. Collects up to five tokens that fit one of my narratives and aren't on the watchlist yet.
Vault ingest. Every saved link gets read and summarized, x.com links through Grok. Max 50 a night tho for cost efficiency reasons.
The Intel page. Signals with a so-what each, Radar blocks, Watchlist table, portfolio movers, attention radar, Vault items, Telegram digest, and an Errors section that lists what it couldn't fetch or verify. Claims carry links. If it couldn't verify something it says so instead of guessing.
The Scripts
Eight scripts plus a tiny key loader, under 1,000 lines of Python. Each one is a thin CLI over one API. Claude calls them as needed or defined.
For X I use Grok. x_scan.py calls the xAI Responses API with the x_search tool, which searches X and returns the posts. Each Radar row has its curated list of accounts, and those handles go into the call so a scan can be limited to the people I actually want to hear from:
Keys are loaded from one keys.env file next to the scripts, uploaded to the Cowork projects as a project file and nowhere else.
Capture
If saving something costs me more than two seconds, I won't do it. And if I do it via bookmarking on X, I probably won't ever revisit it. So, I decided to fix that. In my system, there are six prefixes, and they work everywhere. Any Claude chat, the Telegram bot, an email to myself with the prefix in the subject. Or just tell Claude in chat to "drop X on watchlist" or "add Y to vault".
This is the rule from my global instructions:
On the phone the Notion share sheet goes straight into the Vault. On desktop the web clipper does the same. Forward a post to the bot and it becomes a Vault row with the original author attached.
Typefully
I draft and schedule most of my content in Typefully, so the system had to work around that.
Claude reads the queue for the morning brief: what goes out today and tomorrow, what went out yesterday. When it drafts something, the draft lands in Typefully as a plain draft, tagged "claude" without a date. I open it, edit, maybe loop Claude in again for research support or verifying numbers, schedule, and publish. The Content calendar in Notion mirrors whatever Typefully says and Claude never publishes and never schedules.
Prompts Matter More Than Code
As mentioned earlier on in this article, one reason why I went with Claude Cowork at the center vs. building everything from the ground up is the complexity that comes with the latter and managing/maintaining a huge dev project. Here, me and my bro Claude Code wrote under 1,000 lines of Python, and "just" 5,500 words of prompts and global instructions. So things went much smoother than when I tried to build complex stuff myself in the past.
But guess which above-mentioned half took longer to get right?
Correct, the damn prompts.
The global instructions tell Claude who I am, what I do, what the systems of record are, how Todoist is mapped, what every Gmail label means, the capture prefixes, the tool list and how I want things written, and more. Each project adds its own instructions. Each scheduled run is a prompt with inputs, steps, output format and an Errors section. So it's absolutely crucial to how the system works and performs. Hence took me quite some time.
And sometimes smol fixes are what brings most value-add. For example, one of the instructions that did the most for quality was: "never pad, write 'nothing' for empty sections". Most AI slop is the model filling space. After I removed the space, the briefs read much more like something a real person wrote for me.
Is the system perfect?
No, not yet.
But that's the beauty of it. It runs, and it writes down what it couldn't do. I fix, improve and extend one thing per day. The setup keeps evolving.
What's Next
Improve and extend on front of plain-text queries against the whole thing even more comprehensively and granularly.
Integrate more data/info feeds: DefiLlama, deeper Hyperliquid stats and perp market data, Polymarket, EDGAR, arXiv, etc.
A rating loop where I rate each brief, and the ratings feed back into the prompts to continuously improve the output quality (even tho it's pretty sick already).
Whatever makes sense as I use this thing on a daily basis.
For more sensitive data and tasks where I have some flexibility to model performance, I might also loop in local models running on my own hardware (DGX Spark and Mac Mini). By extending and improving the programmatic logic in the scripts the system has access to, I can also further improve capabilities and outputs as I go.
The only real limit here is ultimately my imagination, especially as @AnthropicAI continues to ship.
Weekend is approaching and I highly recommend to look into this anon.
Can be a real game-changer.